Understanding the Characteristics of Collective Team Movement Based on Feature-Driven Analysis and AbstractionManuel Stein, Johannes Häußler, Dominik Jäckle, Halldór Janetzko, Tobias Schreck and Daniel A. Keim

Visual Soccer Analytics:

CPSC 547 PresentationYann DuboisMarch 14, 2017

ISPRS International Journal of Geo-Information, Special Issue Advances in Spatio-Temporal Data Analysis and Mining, 2015

• Mainly geometrical data

• Data every 100 milliseconds

• Manually annotated events (fouls, goals …)

What: Data

Single Soccer Game

2

Overview

https://www.janetzko.eu/project/soccer/

1

2

1

4

3

Data 5

3

2

• Increasing demand from clubs

• Now we can

• Video analyst: 3 working days per opponent team

• Current support from system is limited

• Visualisation to not get overwhelmed by data

Why

The need of a software

4

• No (good) automatic identification of situations

‣ Need expert verifications

‣ Doesn’t support domain knowledge

‣ : classification method but no explanation

Why

Improve previous work

1

5

• Support experts in exploring characteristics of situations

• Incorporation of meaningful features describing situation

• Visualisation with interactive re-ranking of features and search for similar situations

Why

Tasks

6

Why

7

How

Figure 1. Previous workflow

Workflow

Why

8

How

Figure 3. New workflow

Workflow• Intervals: General time interval • Move: Ball possession• Event: Foul / goal / …

WhyWorkflow

9

How

Interval selection:

• Manual or automatic

• Shows data of interest

• Main reason of use

WhyWorkflow

10

How

• Smooth out noise => better classification

• Less Data

• 100 milliseconds -> 2 seconds time frame

Binning:

WhyWorkflow

11

How

Classification model:• Compute features of binned data

• 5 classification algorithms:

‣ Logistic model trees, Logistic base, Functional trees, decision stump and Support vector machines

• Training set: 33% of intervals

• Returns classified set of 2s intervals

WhyWorkflow

12

How

Game moves and Feature ranking:

• Derive Game moves from interesting 2s intervals

• Extract interpretable features of each moves

• Relevant if unusual values

WhyWorkflow

13

How

Table 1. Meaningful features

WhyWorkflow

14

How

Ranking change:

• User can reranking features

Similarity search:

• Search similar moves based on events and ranking features

WhyVisual design

15

How

Time:• Navigation and Show events

Figure 4. and 5.

Move:• Show moves duration and main feature

WhyVisual design

16

How

Move characteristic:• Shows ranked features

• Connector to see better

• Drag and drop re-ranking

Figure 6.

How

17

Figure 1.

OverviewHow

• 66 professional soccer matches

• Manually annotated events (foul, pass, cross…)

• Temporal resolution: 100 milliseconds

Evaluation

Data

18

• 2 experts : involved in pre-study and expert study

• Coach working at Bayern Munich

• Official referee

• “Ground truth” by additional expert: 35 situations

Evaluation

Expert evaluation

19

Evaluation

Results

20

Table 2. Evaluations results

Evaluation

Results

21

• Experts liked reducing complexity with meaningful features

• Agreed on features

• Proposed to add information on outcome

• Really liked similarity search (and re-ranking)

• Think that video analyst would use it

Discussion

+ strengths

22

• Answer well their task

• Method that you can tweak (reranking) but default => not overwhelming

• Very detailed

• Features seem meaningful

Discussion

- weekness

23

• No video for double check

• Unnecessarily long

• Need to read 1st paper to understand some features

• I would use air / ground and not straightness of ball

Discussion

- weekness

24

• Validation by 2 “experts” but no video analyst

• 66 games dataset in validation but only use 1

• Very important to have a global view of a tactic not precise movement every 2 seconds

• Only single game

• Do not critique their paper

25

Thank you !